AutoTrust @AutoTrustAI

Games are a clear way to see the value of adaptive decision models. GEV-26B-Decide makes a fast System 1 choice first, then brings in System 2 reasoning only when the position is uncertain. In our one-move game tests, adaptive thinking raised Minesweeper from 21% to 86%, Wordle from 52% to 100%, Connect Four from 54% to 99.5%, and Sudoku from 76% to 99.5%. For real-time visual action loops, JEV-27B-VL reads the game state and ranks the next move: Super Mario Bros. movement and jump decisions, Tetris placement, Rubik’s Cube formula selection, Snake steering, and Quick, Draw! recognition. The underlying agent pattern is the same: observe state → rank actions → act → verify → reason more deeply only when needed. GEV-26B-Decide: huggingface.co/autotrust/GEV-… JEV-27B-VL: huggingface.co/autotrust/JEV-…

Adaptive Thinking, Measured. Thank you @HuggingModels for featuring AutoTrust AI’s GEV-26B-Decide. GEV is built for the decision layer inside agent products. System 1 makes a fast, calibrated choice: route a ticket, rank results, select a tool, or choose the next action. When confidence falls below 0.8, it can invoke System 2 reasoning and fold that result back into the final probabilities. The result: simple cases move quickly, while ambiguous cases get more thought instead of forcing a long reasoning trace on every request. In our 1,754-question public evaluation, adaptive mode reached 83.4% accuracy versus 83.8% with reasoning on every question, while calling System 2 for only 47.8% of cases. System 1 takes about 45 ms on one B200. The model card now shows 1.05M downloads in the past month. Download and build with it: huggingface.co/autotrust/GEV-…

Hugging Models @HuggingModels

2 weeks ago

Meet GEV-26B-Decide: a 26B text-classification model that blends system-one intuition with system-two reasoning. It outputs typed decisions with calibrated probabilities, making AI choices more transparent and trustworthy.

3 4 14 2K 5

Thanks for Huggingface community. Now JEV-35B-NVFP4 is open for everyone

Hugging Models @HuggingModels

5 days ago

Meet JEV-35B-NVFP4: a 35B text-classification beast built on Qwen3.5 MoE. It handles image-text-to-text, makes typed decisions, and gives calibrated probabilities. 20k+ downloads already. This is how AI should reason.

3 7 37 3K 22

Thanks for ModelScope

ModelScope @ModelScope2022

6 days ago

JEV-35B turns Qwen3.5-35B-A3B into a fast System 1 decision model, activating only around 3B of its 35B parameters per token. Apache 2.0. 🤖 modelscope.ai/collections/Au… 🏆 Outperforms JEV-27B-VL on the Decision Index public suite while retaining competitive visual decision

keyboard_arrow_left Previous keyboard_arrow_right Next

3 11 95 6K 56

E192 is our most memory-efficient GLM-5.3-Flash build for long-context, tool-using, multimodal workloads. It retains 192 of 288 routed experts through neural architecture search, keeps 18B parameters active per token, and preserves image/video understanding, structured tool calling, and the full 154,880-token vocabulary. For a coding or research agent, that means more usable KV-cache room for working across a large codebase, long technical documents, visual inputs, tool calls, and extended reasoning—not just a smaller checkpoint. The 124 GiB weight footprint is designed for 2× DGX Spark deployments at about 62 GiB per node, or a single 180 GB Blackwell GPU. Its included NVFP4 MTP draft delivered 1.80× faster single-stream decoding on our B200 test with unchanged output quality. On our evaluation runs: 80.0% on AIME 2025, 84.8% on GPQA-Diamond at max reasoning effort, and 88.5% on BFCL v4 non-live tool use. Download the MIT-licensed model: huggingface.co/autotrust/GLM5…

Long Context, More Headroom. Thanks @HuggingModels for featuring AutoTrust AI’s GLM5.3-Flash-E192-DGX-Spark. E192 is our most memory-efficient GLM-5.3-Flash build for multimodal, tool-using agents. It keeps image/video input, structured tool calling, and 18B active parameters per token in a 124 GiB weight footprint. Built for 2× DGX Spark or one B200/GB200. The included NVFP4 MTP draft reached 1.80× faster single-stream decoding in our B200 test, with unchanged output quality. Download: huggingface.co/autotrust/GLM5…

Hugging Models @HuggingModels

5 days ago

Meet GLM5.3-Flash-E192-DGX-Spark: a new image-text-to-text model that's turning heads. Built for blazing fast multimodal tasks, it's already got 4.5k downloads. Curious? Let's break it down.

2 4 14 3K 6

@HuggingModels Thanks @HuggingModels for featuring AutoTrust AI’s GLM5.3-Flash-E192-DGX-Spark. Built for multimodal, tool-using agent workloads with vLLM, NVFP4, and Blackwell-class deployment. @NVIDIAAIDev @huggingface @Zai_org

Thanks for the feature, @HuggingModels. GEV-26B-Decide is for products that must make many small choices reliably: which ticket to route, which search result to elevate, which tool to call, or whether a request needs more investigation. It starts with a fast calibrated decision. When the model is uncertain, adaptive mode can bring in System 2 reasoning rather than spending that extra compute on every request. GEV has now passed 1.05M downloads in Hugging Face’s past-month view. Explore the model, serving scripts, and full evaluation details here: huggingface.co/autotrust/GEV-…

Latest GEV-26B-Decide scorecard: 1.05M downloads in the past month on Hugging Face 62.48 on Decision Index 0.2.1 83.4% adaptive accuracy across 1,754 public questions Only 47.8% of cases required System 2 reasoning About 45 ms per System 1 decision, or 257 decisions/sec with 64 concurrent clients on one B200 The broader AutoTrust AI decision-model family is now showing 3.18M+ downloads across GEV-26B-Decide, JEV-27B-VL, JEV-9B, and JEV-27B in the same rolling-month view. huggingface.co/autotrust

Thanks again for the feature, @HuggingModels. Think of JEV-35B-NVFP4 as a local decision engine for agent products: route an incoming support ticket to billing or technical support; rerank search results for a specific query; choose the right tool for a task; or let a browser agent select its next click from a screen. Instead of producing a paragraph for downstream code to interpret, it can compare up to 256 options and return a calibrated probability for each — so low-confidence cases can be escalated or handed to a human. NVFP4 brings this workflow to a single 32–48 GB GPU: 25.6 GB to download and about 23 GiB for weights. huggingface.co/autotrust/JEV-…

Thanks again for the feature, @HuggingModels. JEV-35B is built for the decision layer inside real agent workflows: it can rank candidates, route requests, select tools, or choose the next computer-use action from text, structured state, or images — evaluating up to 256 options in one pass and returning a calibrated probability for each. Built on a 35B MoE model with ~3B active parameters per token, it reached 59.73 on our Decision Index 0.3 public-suite run. We also tested 95% completion across 60 multi-step browser tasks with element text available. Weights, vLLM serving script, and `/v1/decide` API: huggingface.co/autotrust/JEV-…

What can you build with it? • A coding agent that reads a large codebase, inspects screenshots or error logs, calls tools, and works through a fix with configurable reasoning effort. • A multimodal research assistant that reads long technical material alongside figures, charts, and images — then synthesizes evidence, writes analysis, or prepares structured outputs. • A document and visual-data workflow: inspect an infographic, chart, scanned report, or product image, then extract what matters and turn it into an actionable answer. • A tool-using application that switches between fast everyday requests and deeper reasoning for difficult coding, science, math, or planning tasks. E256 supports image and video inputs, structured tool calling, and long-context reasoning. For interactive serving, its included NVFP4 MTP draft delivered 1.84× single-stream decode speed in our B200 test. Model card and deployment details: huggingface.co/autotrust/GLM5…

Thank you @HuggingModels for featuring AutoTrust AI’s GLM5.3-Flash-E256-DGX-Spark. This is our largest Blackwell-oriented GLM-5.3-Flash build: 145B total parameters, 18B active per token, with 256 of 288 routed experts retained through neural architecture search. It keeps multimodal image and video input, full vocabulary support, structured tool calling, and long-context reasoning. NVFP4 experts and an FP8 KV cache cut the weight footprint to 170.5 GB. The included NVFP4 MTP draft reaches 1.84× single-stream decode speed on a B200 with unchanged output quality. On our evaluation runs, E256 reached 97.6% on HumanEval, 89.4% on C-Eval, and 76.1% on MMMU. Download the MIT-licensed model and run it with vLLM: huggingface.co/autotrust/GLM5…

Hugging Models @HuggingModels

5 days ago

Meet GLM5.3-Flash-E256-DGX-Spark: a multimodal MoE model that takes image+text in and generates text out. Built for vLLM and optimized with NVFP4 on Blackwell GB10, it's a lean, fast beast for next-gen AI apps. 3k+ downloads already!

4 7 24 2K 9

JEV-35B-NVFP4 for fast, private product decisions: • Support/review routing • Search & recommendation re-ranking • Browser agents: screenshot → next click • Visual QA 25.6GB download, 23.3GiB vLLM memory—runs on a 32–48GB GPU. Same /v1/decide API. Download: huggingface.co/autotrust/JEV-…

Thanks @HuggingModels for featuring AutoTrust AI’s JEV-35B-NVFP4. This is an open Apache-2.0 System 1 decision model for building real agent loops: it reads text, structured state, or images; evaluates up to 256 options in one pass; and returns a calibrated probability for every action through `/v1/decide`. The NVFP4 release reduces the weights from 71.9 GB to 25.6 GB and GPU memory to 23.3 GiB — enabling local deployment on a single 32 GB+ GPU. It preserves 96.6% of BF16 decision outputs across 308,559 Decision Index answers. Already 33K+ downloads. Download it, run the included vLLM script, and build with it: huggingface.co/autotrust/JEV-…

Hugging Models @HuggingModels

5 days ago

Meet JEV-35B-NVFP4: a 35B text-classification beast built on Qwen3.5 MoE. It handles image-text-to-text, makes typed decisions, and gives calibrated probabilities. 20k+ downloads already. This is how AI should reason.

3 7 37 3K 22

Built for product decisions: • Route tickets/tasks with thresholds • Re-rank search, recommendations, review queues • Drive agents: screenshot → click → updated state 204 ms median /v1/decide (B200); 95% on 60 browser tasks with element text. Download: huggingface.co/autotrust/JEV-…

Open System 1 Decision Models Thank you @HuggingModels for featuring AutoTrust AI’s JEV-35B. Built on Qwen3.5-35B-A3B, JEV-35B activates about 3B parameters per token and is designed for fast, confidence-aware decisions inside agent workflows. It accepts text, structured state, or images; compares up to 256 options in one pass; and returns a calibrated probability for every action through `/v1/decide`. On our Decision Index 0.3 public-suite run, it scored 59.73, ahead of JEV-27B-VL’s 53.64 public score. It also achieved 51.3% on ERQA in one pass with no robot or embodied data used in decision training. Download the Apache-2.0 weights, run the included vLLM server, and build with it: huggingface.co/autotrust/JEV-…

Hugging Models @HuggingModels

5 days ago

Meet JEV-35B: a 35B parameter MoE model that doesn't just classify, it makes typed decisions with calibrated probabilities. Built on qwen3_5_moe, it's the first open model designed for system-one thinking. 17k downloads and counting.

4 10 57 3K 36

NVFP4 makes JEV easier to deploy: • Only 10,240 routed expert MLPs are NVFP4; everything else stays bf16 • 71.9→25.6GB download; 66.5→23.3GiB vLLM memory (−65%) • 96.6% same top-1 decisions vs bf16 across 308,559 answers Same /v1/decide API. Download: huggingface.co/autotrust/JEV-…

@ModelScope2022 There is also a live AutoTrust AI demo: news recommendation and search reranking powered by calibrated System 1 decisions. Try it here: news.guru.so/ui Model card: huggingface.co/autotrust/JEV-…

JEV-35B is now on ModelScope. Thank you @ModelScope2022 for featuring AutoTrust AI’s latest decision model. JEV-35B turns Qwen3.5-35B-A3B into a calibrated System 1 decision model: • 35B total parameters, with ~3B active per token • Selects from up to 256 text or image options in one pass • Returns calibrated probabilities through `/v1/decide` • 59.73 on the Decision Index 0.3 public suite • 204 ms median decision latency on one B200 • Apache-2.0 For local deployment, the NVFP4 version is a 25.6 GB download and uses about 23 GiB of GPU memory, making it practical on a single 32–48 GB GPU. Try it on ModelScope: modelscope.ai/collections/Au… Model card and vLLM quick start: huggingface.co/autotrust/JEV-…

ModelScope @ModelScope2022

6 days ago

JEV-35B turns Qwen3.5-35B-A3B into a fast System 1 decision model, activating only around 3B of its 35B parameters per token. Apache 2.0. 🤖 modelscope.ai/collections/Au… 🏆 Outperforms JEV-27B-VL on the Decision Index public suite while retaining competitive visual decision

keyboard_arrow_left Previous keyboard_arrow_right Next

3 11 95 6K 56

意思決定向けマルチモーダルモデル「JEV-35B」のNVFP4量子化版が公開。 全10,240基のエキスパートを量子化し、GPUメモリ消費を約65%削減。 bf16版の判断・視覚精度をほぼ維持し、推論の省リソース化に寄与。 #JEV35B #vLLM URLはリプ⬇️